How does Netflix know what we want to watch?

By Karl Quinn, May 22, 2022. The Sydney Morning Herald

Choosing what to view from hundreds of options is hard work for humans. So, Netflix curates your options
for you. How? And should you be grateful or alarmed? Netflix has become such an integral part of our
lives, it’s easy to forget it officially arrived in Australia only in March 2015. Although the company doesn’t
reveal precise numbers, audience research suggests there are more than 5 million subscriptions in Australia,
with close to half of all Australians having access to the service.

The market has become crowded with competitors such as Amazon’s Prime Video, Apple TV+, Disney+,
and Paramount+, but Netflix is still the market leader by some margin. And it’s not just what Netflix offers,
but how it offers it, that has put it there. To many “members”, as it calls its customers, Netflix has become
a trusted friend – always there, always with something new to watch and always, always, willing to
recommend a show that might just tickle their fancy. Willingness to splurge on programming – the company
is expected to spend close to $US18 billion ($26 billion) on content this year – plus the ability of viewers to
binge-watch entire series in one sitting, and the lack of advertising (at least until the end of this year, when
a cheaper ad-supported option is likely to be introduced) have been critical factors in the platform’s success.

But so is its ability to siphon the more than 7000 movies, documentaries and series available on the
Australian service through a funnel that seems specifically tailored for each and every one of us, to make
viewing suggestions that, more often than not, are at least in the ballpark of what we might actually want
to watch at that moment. That funnel is what’s colloquially known as “the Netflix algorithm”, and it kicks
in when you log in and answer the question: “Who’s watching Netflix?” But what is this algorithm? Should
you be scared of it? And does it actually help you pick the best of what’s on offer, or hinder you?

What is this ‘Netflix algorithm’?

An algorithm is simply a set of rules or instructions to be followed in a calculation – in this case, the
calculation that tries to match program suggestions to a member’s tastes with the highest degree of accuracy.

In the case of the so-called Netflix algorithm – the company prefers the term “recommendation system” –
the calculation is performed by automated programs, not humans, and each time it makes a suggestion and
you respond by selecting or not selecting that recommendation, by watching or not watching that program
all the way through, the algorithm gets a little smarter. It’s called machine learning, and yes, being just a
little bit scared (especially if you’ve ever watched a Terminator movie) is not unreasonable. Our robot
overlords are indeed watching us – or at least watching what we watch.

You can see the Netflix algorithm in action on the home page on your screen, where row after row of
categorised recommendations appear, all of them tailored to your user profile. If you’re in a household with
multiple users, try logging on with someone else’s profile, and you’ll notice their home page probably looks
quite different to yours. It’s not just that different titles are recommended – that much you’d expect if
you’re an adult, say, and they’re a kid – but there’s probably a very different ordering of the rows within
which those titles are served up. So, when one member logs on, they might see the Because You Watched
row up top, but another member might get something entirely different – Top 10 in Australia, say, or New
Releases or Historical TV dramas. And it’s not just the order of the rows, either – it’s the rows themselves,
which can also differ widely from user to user. That’s because there isn’t just one Netflix algorithm, there’s
a whole bunch of them. Each member gets a home page tailored to their “likes” (as indicated by the thumbs-
up or thumbs-down rating they might assign after watching a show or movie), their viewing history and
similarities between their profile and other members’, across the entire world. It’s almost as if Netflix
actually knows you.

How do the algorithms work?

It all starts with that user profile you create. This is the anchor to which all the viewing data you generate
is tethered. Every time you watch something while logged in under your profile, Netflix feeds that
information into its database in a bid to predict what you might like to watch in future. If you start watching
something but don’t finish it, it records that. If you binge a 10-part series in a single day, it takes note of
that, too. Perhaps the most powerful tool, though, is the thumbs-up/down system, which in 2017 replaced
the five-star rating system Netflix had used for 20 years, all the way from its beginnings as a mail-order
DVD service. In April, Netflix tweaked it again, introducing a double thumbs-up option because, it claimed,
“we’ve learned over time that these feelings can go beyond a simple like or dislike”.

The old star system was a ranking of what people thought of a title after watching it. In 2006, Netflix began
trying to predict what rating a title was likely to receive before people had rated it, so as to better recommend
titles to them. This was the beginning, in very rudimentary and complicated form – an early version utilised
107 different algorithms to produce a single result – of the recommendation system in place today. From
the simplified ratings system now in place, Netflix claims, “We get a better idea of what you’d like to watch
by looking at things such as the genres of TV shows and movies, your previous ratings, your viewing history,
and the ratings of Netflix members who have similar tastes to you”. But unlike the old star system, the
percentage figure is not a rating per se, nor even a prediction of how you will rate it. “The percentage next
to a title shows how close we think the match is for your specific profile,” Netflix says.

Why does the recommendation system even exist?

Two reasons, really: to serve the business, and to serve us. Netflix’s business isn’t built, like traditional
television, on ratings and selling advertising (for now, at least). It’s all about subscribers: acquiring new
ones, maintaining existing ones and winning back lapsed ones – an even more pressing part of the equation
now that the company’s subscriber numbers have begun to go backwards for the first time in a decade. The
higher the accuracy of recommendations, the less reason a customer has to drift away or question the value
of their subscription. And for members, it’s all about making the path from option to decision as smooth,
simple and swift as possible. Consumer research suggests that a typical Netflix member loses interest after
perhaps 60 to 90 seconds of choosing. “Humans are surprisingly bad at choosing between many options,
quickly getting overwhelmed and choosing ‘none of the above’ or making poor choices,” a couple of Netflix
developers wrote in a technical paper published in 2015. “Consumer research suggests that a typical Netflix
member loses interest after perhaps 60 to 90 seconds of choosing, having reviewed 10 to 20 titles (perhaps
three in detail) on one or two screens [rows].”

In other words, we’re easily overwhelmed by too much choice, so the job of the recommendation system
is to make it easy for us by narrowing the options. But merely restricting choice won’t necessarily produce
a great outcome. To keep members happy, the recommendations need to be both manageable in number
and right for an individual more often than not. If you select a show that carries a “96 per cent match”
score, you start watching it, and you hate it, that counts as a fail. The system can’t wear too many of those
or it risks creating dissatisfied customers, who might quickly become former subscribers – so the algorithm
factors in your non-completion of the show and adjusts the next recommendation accordingly.

Netflix’s product-development teams are continually working on minor tweaks to make the
recommendation system simpler, everything from moving billboards on screen and short video “synopses”
to tailored images for titles (you and I might see different images for the same show, based on our previous
viewing history and preferences). They continually carry out “A/B testing”, whereby different options are
presented to a percentage of the global audience to see which produces the better outcome. That then
becomes part of the relevant algorithm. It’s all designed to maximise what Netflix calls the “take-rate” –
the rate of conversion from recommendation to watch. But all this data doesn’t just inform what gets
recommended to us, of course. It also helps guide Netflix in deciding which shows to buy for the service,
which ones to keep, which genres to back and which actors, writers and directors to invest in.

It is data that has driven Netflix’s push into Korean content, a push validated by the enormous success of
Squid Game, the service’s most-watched title. It was data that drove Netflix to sign a four-picture $US250
million deal with Adam Sandler in 2014, and to extend that deal in January 2020 for another four movies.
“Whether you know him as Sandman, the Water Boy, Billy Madison, Happy Gilmore, Nick Spitz or simply
Adam, one thing is clear: our members can’t get enough of him,” Netflix chief content officer Ted Sarandos
said of Sandler in announcing the extension. “They love his stories and his humour.”

Based on viewing hours in the first 28 days of the title on the service. The top 10 are all series; the most-
watched movie so far is Red Notice with about 364 million hours. Netflix’s subscribers love Friends, too,
which is why the company paid $US100 million for the series in 2015, despite the fact it had been in rerun
on free-to-air television for years. By some estimates, it was the platform’s second most-watched show
(after the US version of The Office). But when Netflix lost Friends in 2019 to Warner Media – which was
reclaiming its back catalogue ahead of the launch of its own streaming service, HBO Max, in May 2020 –
it spent a whopping $US500 million to shore up the rights to Seinfeld in its place. It was a clear demonstration
of the immense value of good content with a proven audience to a business built on retaining subscribers.
Those are big numbers, but they are all about giving people what the data says they want, and giving them
plenty of reasons to remain as subscribers.

With everything determined by past behaviour, am I ever going to stumble on something great again? That’s
what the shuffle feature, introduced in September 2020, aims for, but really the answer is probably not, at
least not in the way you might have back in the days of browsing the shelves of the local video store
(remember them?). The more complex answer is that Netflix is aware that the lack of randomness, of the
happy surprise, is a limitation, and it is trying to factor in ways to compensate for this. But how exactly do
you program the unexpected? One solution is the Top 10 and Now Trending rows, which propose titles
users might not otherwise have discovered if left to their own devices.

The trouble with this is it’s dictated by the lowest common denominator: a title might be popular not
because it’s good but purely because many other people have watched it, for any one of a variety of reasons
(social media chatter, perhaps, or accidental newsworthiness). In effect, that creates a feedback loop: a show
becomes popular for some reason unconnected to its quality, it shows up in the most popular row, curious
viewers check it out, and that reinforces its popularity. How else to explain the sudden popularity of the
much-derided escapist fantasy of Emily in Paris during the depths of lockdown?

The dedicated explorer can, of course, escape the tyranny of recommendations by using the Movies or TV
Shows buttons at the top of the screen and selecting from the pull-down menus. If all else fails you can
simply leave your viewing fate entirely with the Netflix gods by using the Play Something button. Select
LGBTQ under Movies, for example, and you’ll get a landing page with a bunch of new rows, including,
perhaps, Documentaries, Romance, European Movies, Asian Movies, Independent Movies and Netflix
Originals. If you’re a fan of Sylvester Stallone action films, chances are none of those would have shown
up on your home page. But now you’ve done this, of course, they just might. For a serious shot at
randomness, you can even type the word “random” in the search bar. Trust us, you’ll get exactly what it
says on the box. If all else fails, you can simply leave your viewing fate entirely with the Netflix gods by
using the Play Something button. Introduced in April 2021, originally with up to 75 titles, the feature now
presents users with about 300 options, seemingly random but actually curated using your viewing history
and likes as a guide.

This is pretty random: I once watched Austin Powers and it recommended The Boy in the Striped Pyjamas. Why?
True story, this. It’s impossible to say with certainty why a comedy about a shagadelic ’60s spy might lead
to a harrowing tale about Nazi concentration camps, but it’s unlikely Netflix thinks the Holocaust was a
subject for humour. It is just about conceivable, though, that the confusion was sparked by something as
simple as an image of Austin Powers star Mike Myers in a striped Carnaby Street suit (striped suit, striped
pyjamas) because images have become such a crucial factor in all this. Since May 2016, different users have
been served different images to promote the same title – with those images chosen on the basis of what
Netflix knows about them. The reason? Most of us make our viewing choices based on the image, and we
do so very, very quickly.
“Artwork was not only the biggest influencer to a member’s decision to watch content, but it also
constituted over 82 per cent of their focus while browsing.” Announcing this new approach, Netflix noted
“artwork was not only the biggest influencer to a member’s decision to watch content, but it also constituted
more than 82 per cent of their focus while browsing Netflix”. It added, “users spent an average of 1.8
seconds considering each title”. That’s not a lot of time to delve deeply into plot details, but enough to
make a snap image-based judgment about whether to hit play or not. Because it is impossible to know
precisely which images any one of us will see, we’ll take a hypothetical example. Let’s say the Australian-
made, post-apocalyptic, sci-fi thriller, I Am Mother, turns up in your feed because you have watched a lot of
sci-fi on Netflix lately. Chances are you’ll see an image of the movie’s robot, Mother (lovingly built by those
special effects wizards at Weta in New Zealand, the same people who gave us all those hobbits and orcs in
Peter Jackson’s Tolkien movies). But if you’ve just watched and given the thumbs-up to Million Dollar Baby,
you might be served an image of Hilary Swank instead (she’s one of the stars of I Am Mother, and won the
best actress Oscar for Clint Eastwood’s 2004 boxing movie). It’s ingenious, but it can have unintended
outcomes, as some African-American customers noted in 2018 when they were served title cards featuring
black actors who had only minor roles in shows. An embarrassed Netflix was forced to once again declare
it did not use ethnicity as a filter – presumably, it was viewing history alone that produced that result? – but
people quite rightly wondered precisely what information they were handing over every time they declared
who was watching.

Do I need to play nice with Netflix on all this?

It’s up to you whether you want to be an active participant in the whole thumbs-up/thumbs down thing,
but even if you don’t, you should know that Netflix is harvesting an awful lot of data on your viewing
behaviour anyway. The company swears it does not share this information externally but, as we’ve
discussed, it does use it to determine what sort of programs it commissions and who it gets to direct them
and act in them, as well as what it serves up to you. So, frankly, unless you just don’t trust it with your data,
you might as well get in there because chances are it’s going to improve your experience of the service and
others like it (Prime Video and Stan, in particular). If you watch a show and don’t rate it, a crucial piece of
data is missing from the feedback system. The recommendation system is powered by machine learning
but the algorithms that drive it are only as good as the inputs they receive; garbage in, garbage out, as they
say in the computing world. If you watch a show and don’t rate it, a crucial piece of data is missing from
the feedback system (although if you watched only the first seven minutes, say, of a two-hour movie, that
still tells the AI something).

Cite this article:

Quinn, K., (2022). How does Netflix know what we want to watch? The Sydney Morning Herald.
https://www.smh.com.au/culture/tv-and-radio/how-does-netflix-know-what-we-want-to-watch-
20220512-p5akmf.html